Beyond Accuracy: A Decision-Theoretic Framework for Allocation-Aware Healthcare AI

Artificial intelligence (AI) systems increasingly achieve expert-level predictive accuracy across healthcare tasks, yet their real-world clinical impact remains inconsistent. Prior studies demonstrate that improvements in model performance often fail to translate into improved patient outcomes-a disconnect this paper terms the allocation gap. We provide a decision-theoretic explanation for this phenomenon by formalising healthcare delivery as a scarcity-constrained stochastic allocation problem. Within this framework, AI functions not as an autonomous decision-maker but as decision infrastructure that improves outcomes through utility estimation, variance reduction and information prioritisation. We employ formal tools from constrained optimisation and Markov decision processes (MDPs) to analyse how improved estimation alters optimal resource allocation under binding capacity constraints. A controlled synthetic triage simulation demonstrates that allocation-aware policies, which explicitly optimise for constrained utility, outperform conventional risk threshold policies by 18-25% in realised utility, even when predictive accuracy is identical. The framework provides a principled basis for evaluating, deploying and governing healthcare AI beyond task-level metrics, with direct implications for resource-constrained clinical environments.